Traing KITTI result

Question

Traing KITTI result

CuongNguyen218 opened this issue 5 years ago · 6 comments

Hi @jwchoi384 ,
After training process with 60k iteration, I got this result. I see a big gap between result from your paper and my result in Car Moderate mAP.And I want to hear your idea about the gap? Is it normal ?

Answer 1 · 2019-12-16T14:26:29.000Z

Can you provide train_kitti_list and test_kitti_list? i realize that my test_list have 21k box and yours have 17k box.

Answer 2 · 2019-12-17T08:30:22.000Z

@CuongNguyen218
Hi, I uploaded KITTI train/validation list in my repo.
train_kitti_3cls_list
valid_kitti_3cls_list
As I mentioned in our paper, the KITTI training and validation sets are made by randomly splitting the training set in half.

Answer 3 · 2019-12-17T08:39:39.000Z

I want to hear your advice about my result @jwchoi384

Answer 4 · 2019-12-18T14:27:26.000Z

@CuongNguyen218
I do not know the exact reason. It seems to be due to random training (ex. augmentation, resize, initialization etc.) or difference of training/validation list maybe.
Anyway, your results seems good.

Answer 5 · 2019-12-23T08:37:24.000Z

@jwchoi384 ,

I used your train-val spit and get this result on validation set at resolution 512 x 512 for car class. When i was training, I remove labels like tram, van, truck out of my dataset. How do you think about my result ? I realize that my result for car class is higher than what you get in your paper.

Answer 6 · 2019-12-23T08:54:05.000Z

@CuongNguyen218
I also removed labels like tram, van, truck when training and validation.
In my opinion, performance may vary depending on random initialization, data augmentation, and system environment.